An optimal opinion tree generation method for aspect sentiment quad prediction
Enhancing tree structure representations using contrast learning is a mainstream approach to applying tree structures to aspect based sentiment analysis(ABSA),but the method inherently relies on data enhancement,which may remove semantic information relevant to downstream predictions and makes it difficult to obtain optimal structures.To address the above issues,an optimal opinion tree generation method with structural entropy constraints for aspect sentiment quad prediction(ASQP)is proposed,which is implemented by a text encoder and a structural encoder that directly generates positive samples.The structural encoder takes the opinion tree embeddings as input,extracts the intrinsic essential informa-tion inherent in the opinion tree using the structural entropy minimization principle,and injects this infor-mation into the textual representation through representation learning.Experiments on two common data-sets validate the superiority of the method.
万方数据
维普
